*****

Patrick Williamson is an English poet and translator, and part-time lecturer on a master’s in translations at ESIT, Sorbonne-Nouvelle Paris 3. Latest poetry collection: Presence/Presenza (English-Italian, Samuele Editore, 2023). Editor and translator of The Parley Tree, Poets from French-speaking Africa and the Arab World (Arc Publications, 2012) and translator notably of Max Alhau, Tahar Bekri, Guido Cupani and Erri de Luca. Member of transnational literary agency Linguafranca and the European board of The Antonym.

Patrick is no stranger to The High Window and in June 2022 he curated a selection of Francophone poetry from Africa and the Arab World which you can access bby following the link.

*****

Poetry and machine translation

Those familiar with automatic translation know very well that the key benefit is to provide assistance with translation with a sort of toolbox that makes the process more efficient. The quality of output text is improving and engineers can create specific software for doctors or lawyers, for example, as there are specialised terms that are easy to encode. The whole difficulty in the literary field is that there are no specialities, not only are we independent of the other’s culture, but what we are dealing with is chances and semantics that are extremely variable. That said, I contend that it is a useful tool when translating poetry. Translation requires the mechanical action of typing, and a version extracted from a machine can provide a canvas for the human translator to work on, using the same cognitive processes as for other translation. Notably, it leaves the translator more time to check, mull and review, especially when working to a tight deadline.

There has been research work done on poetic machine translation, mostly concluding it has its limitations. By way of illustration, note Genzel et al who acknowledge that “Machine translation of poetry is probably one of the hardest possible tasks that can be considered in computational linguistics, MT or even AI in general”. However, they also think the problem of the poetic form can be tackled via producing translations with meter and rhyme for phrase-based MT. At the other end of the ‘spectrum’ Baumann et al apply neural networks to address the challenge of translating poems that do not use any punctuation, as found very often in modern and post-modern poetry. In all cases though it is clear, as Brad Girardeau and Pranav Rajpurkar state, that: “Poetry is particularly difficult to translate, as meaning and form are intricately interwoven. […]. machine translation systems continue to improve at translating a variety of texts. However, these systems focus on preserving meaning, not capturing the rhythm and sound qualities important in poetry.” Lastly, Kurt Gödel’s incompleteness theorems have forced mathematicians to accept certain limitations of their discipline and this could apply to machine translation software developers as there is ‘no consistent system of syntax rules whose poems can be listed by an effective procedure (i.e., an algorithm) capable of generating all target poems”. In sum, there will probably never be a computer program able to achieve a full and exact poetic translation, or translate all poems. Even the most advanced models are limited to a small number of poetic and stanza structures.

In general, the translator rarely succeeds in keeping the entire set of features of the original poem. That is because the same poem, when read by different people of different cultures may trigger completely different meanings and emotions. There is a multidimensional understanding of the music of the words combined with their meaning and the concurrent excitation of emotion. When a human translator takes in a poem in the source language, they first groke the entire set of features conveyed by the poem. They must deeply understand the nuances, cultural references, word fusions if any, textual structure, rhyme and tempo, and many other features unique to that specific poet and poem. Only once the translator has internalized the full set of interplaying features, does he or she have a comprehensive mental model of the poem. The target poem must convey all the features of the model and their relationships, as represented in the translator’s unique brain neurology. Translations of the same source-poem done by different translators are thus all singular newly-created poems, each one an artistic utterance of its own.

It is worth noting that the human translator, proficient in both source and target languages, translates a poem by employing the same process as used in all forms of what is commonly perceived as translation. Translators engage in mental stylistic, structural and content editing as they work in order to tailor the target language. Donald Kiraly’s model of the cognitive processes involved in translation also sets out that translators have a relatively uncontrolled processing centre (intuitive, less conscious) and a controlled processing centre (strategic, more conscious). The former is a key feature for translators of poetry in view of the multidimensional nature of a poem. It is often the case of the mot juste emerging from an unknown origin. The latter considers any translation problems brought to light in the intuitive workspace and implements a strategy to deal with them.

Machine translation is in essence such an uncontrolled processing centre. It generates words or expressions that create material that might not have emerged from the translator’s own intuitive centre. These can then be discarded, kept, or amended depending on the sensibility of the poet and obviously the author’s meaning and intention. The software may yield new ideas that refresh the translator’s approach and options along the lines of “that sounds good”, “why not”, “that gives another angle to consider”. It acts as an external memory, in the same way as poet-translators may refer to print dictionaries or thesauruses to check meanings and to stimulate the word choice/direction of the translation. The same can be said for when the translator engages in collaborative approaches with revisers and/or the author, where a back-and-forth in maieutic style draws the answers out from their memories.

The translator applies the same rules as for translation in general during the reviewing and editing stage: reread, check, reformulate, apply critical thinking, analyses of tone and register, research into the background for possible references, and, depending on the case, formal style. In sum, a whole range of tasks that form part of the methodological work of the translator. This action equates to the ‘controlled processing centre’ but clearly does not preclude the translator’s intuitive side from contributing. Lastly, the translator may use another software tool to extract a target text with variations that may confirm or invalidate the proposals generated by the initial tool, and deliver other options the human translator can use.

Moreover, machine translation, even in its infancy with regard to its capacity for poetry translation, may also serve a purpose in terms of co-translation – a practice which involves translating from a language you cannot speak. The way this works is that two people make the translation, one who is fluent in the source language, and one who is fluent in the target language. One well-known example of is Ted Hughes’ translations of poems by János Pilinszky. In Hughes’s case, for instance, he didn’t speak any Hungarian. Instead, he worked alongside Hungarian writer János Csokits, who prepared for him careful word-for-word English versions of Pilinszky’s poems. Hughes would read these versions, and re-translate them into better English in order to make something that looked like a finished English poem. Ted Hughes wanted to grasp the spirit of these poems rather than recreating the rhymes and tight patterns of Pilinszky’s originals. Indeed, I have also used machine translation to translate a poem by Guido Cupani from Italian, a language I do not master, whereby I revised the output text according to my poetic individuality. I subsequently sent my initial translation to Guido who, being fluent in English, revised both content and format such that the two poems are individual works. In these cases, it is the creative spirit, insight and skill of both parties that prevails, creating a new work entirely.

In sum, machine translation software is merely based on a series of linguistic or statistical rules and encoding of the source and target-language words. It lacks the intuition of the human brain with regard to the exact meaning of a word depending on the context and its established or new relationship with the other features of the poem. The human factor remains key throughout to ensure that the intentions of the original poem are respected. In the current state of the technology, human intervention is not secondary but crucial and a form of creation. The machine itself is not capable of executing a work, it merely extracts the substance of texts and makes this available to translators. As Claire Larsonneur has remarked: “The machine is capable of creating something quite paralysing in the end, so the translator has to work on ‘restoring a richness that otherwise made lost in the automatic translation’”. These machines are complementary tools that are becoming more effective but they are still tools.

References

Baumann, T., Meyer-Sickendiek, B., Hussein, H. (2020). How to identify speech when translating unpunctuated poetry. Conference paper

Genzel, D., Uszkoreit, J., Och, F. (2010) “Poetic” Statistical Machine Translation: Rhyme and Meter, Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, MIT, Massachusetts, US

Girardeau, B., Rajpurkar, P. (2013) Poetic Machine Translation

Kiraly, D. (1995). Pathways to Translation Pedagogy and Process. Kent, Ohio, US: Kent State University Press

Larsonneur, C. (2022). Contribution to the roundtable on automatic translation, Littérature à l’épreuve de sa transculturation Acteurs, espaces, relais, manifestations, Les Lilas, France

Mossop, B., Hong, J., Teixeira, C. (2020). Revising and Editing for Translators. Fourth edition. Routledge.

Sadeh, D., (2019) retrieved at Poetry and the Machine Translation Problem | by Doron Sadeh | Medium on 23 August 2023

Back to the top